Bandwidth estimations

Size: px

Start display at page:

Download "Bandwidth estimations"

Blake Carpenter
6 years ago
Views:

1 Bandwidth estimations for Test Traffic Measurements project Henk Uijterwaal Mark Santcroos RIPE NCC New Projects Group BEst2003, December 9 1

2 RIPE NCC Regional Internet Registry for Europe Based in Amsterdam, The Netherlands New Projects group 2

3 Test Traffic Measurements Worldwide network of measurement nodes (however focussed on Europe) Active end-to-end measurements Measuring performance of Internet infrastructure: Packet Delay Packet Loss Jitter (Packet Delay Variation) Routing vectors (traceroute) DNS servers (root, gtld's & cctld's) GPS synchronized 3

4 Test Traffic Measurements (2) Typical hosts of Testbox: ISP's Internet Exchanges Universities Network research institutions Not doing measurements for the sake of doing measurements Provide (end) users of testboxes with understandable data TTM -> Information Services 4

5 Bandwidth measurements Want to add bandwidth measurements to our list of measurements: TB hosts can check what they are paying for TB hosts can plan network upgrades Want to measure: Capacity bandwidth Available bandwidth Use (and / or improve) existing tools, no need/time to write another one 5

6 BW requirements for TTM Non-intrusive for the network Reliable and consistent Not interfere with other measurements on the Testboxes Relatively fast 6

7 Traffic light Creating a green, orange and red light is simple of course Determining when to show what color is the problem For delay's we have alarms, for bandwidth we could do the same 7

8 What we did... First idea (some time ago): Write some wrappers around existing tools Ran it in production network Too many problems Second idea (more recently): Test the tools in a controlled testbed Very simple parameters Results still not satisfying 8

9 Result Only more questions, but I'll come to that later... 9

10 Test environment Two machines connected with cross-cable 100 Mbps full-duplex tt95: PII 333 MHz, 64 MB, 3Com NIC, FreeBSD 4.6 tt96: PII 466 MHz, 256 MB, 3Com NIC, FreeBSD

11 Test environment (2) Leaping ahead of conclusions: probably too slow 11

12 Throughput tools Netperf and Iperf Saturate the link with UDP and TCP TYPE FROM TO Payload (Mbps) Ethernet (Mbps) UDP tt95 tt UDP tt96 tt TCP tt95 tt TCP tt96 tt

13 New tool Capacity Tool From To Capacity(Mbps) pathrate tt95 tt96 92/93 pathrate tt96 tt95 88/89 13

14 New tools Available BW (no cross-traffic!) Tool From To Capacity (Mbps) pathload tt95 tt96 100/111 pathload tt96 tt95 96/97 pathchirp tt95 tt96 67 pathchirp tt96 tt

15 CPU influence How does the CPU frequency have influence on the results? Results from all tools different in the two directions between tt95 and tt96 Find out what is the real influencing factor! 15

16 Motivation Evaluation of tools showed there are more parameters than just bandwidth Doing Network Delay Quantification study Try to find out how delay is build up on it s way (Will also help understand network delay distribution for other TTM measurements) 16

17 DAG setup TT95 TT96 BIJLMER w. DAG 17

18 DAG Card 18

19 DAG Card Xilinx FGPA equipped NIC realtime hardware timestamps GPS synchronized First 64 byte of frame 19

20 Timestamps: where? SOFTWARE USERLAND KERNEL RECV BPF HARDWARE ETHERNET DAG Mean Std dev Median 95% Max DAG 95 us 22 us 89 us 148 us 225 us BPF 194 us 23 us 190 us 249 us 332 us RECV 405 us 43 us 401 us 468 us 911 us 20

21 Delay layout End-to-End delay is build up from several components D1 = SEND <->DAG D2 = DAG <-> BPF D3 = BPF <-> RECV Mean Std dev Median 95% Max D1 95 us 22 us 89 us 148 us 225 us D2 100 us 4 us 99 us 107 us 114 us D3 211 us 29 us 207 us 249 us 600 us 21

22 Clock accuracy What is the error of various clocks involved? Trimble antenna accuracy: ~50ns to UTC System + NTP + GPS: ~1us to UTC DAG card + GPS: ~120ns to UTC Still small enough with today s delays 22

23 NIC influence How do NIC s work? Could be of influence Will try different NIC s 23

24 Status summary Promised this as a service a while ago Results still not satisfying enough We still want to deliver at some point Doing more research on all related matters 24

25 Plans More careful logging of all results Delay measurements at different rates Analyse real traffic to get (more) insight about available bandwidth properties Investigate ways to improve sending process 25

26 Questions and discussion 26

Test-Traffic Project Status and Plans

Test-Traffic Project Status and Plans Henk Uijterwaal, Fotis Georgatos, Johann Gutauer, Daniel Karrenberg, René Wilhelm RIPE-NCC New Projects Group RIPE-36, Budapest, May 2000 1 Overview Manpower, Email